Abstract
Autonomous drone racing represents a significant challenge in robotics, requiring high-speed navigation through complex 3D environments under strict real-time constraints. While Iterative Learning Model Predictive Control (LMPC) has proven effective in autonomous car racing, its application to aerial platforms is often hindered by 3D dynamical complexities and the risk of aggressive corner-cutting. This paper introduces a novel LMPC framework tailored for drones, incorporating a Cartesian-based formulation, an adaptive cost function, and a shifted local safe set. Experimental results demonstrate that this approach enables drones to iteratively refine their trajectories, achieving a 60.85% improvement over baseline controllers and a 6.05% improvement over state-of-the-art racing algorithms while maintaining real-time performance.
Introduction
Autonomous drone racing pushes the boundaries of control theory and trajectory optimization. To achieve peak performance, a controller must navigate a series of gates at the physical limits of the vehicle's dynamics. Currently, the field is divided between Reinforcement Learning (RL) and Model-Based Control. RL approaches often suffer from sample inefficiency and poor sim-to-real transfer, while Model-Based methods, though safer, are notoriously difficult to tune for absolute time-optimality.
Iterative Learning MPC (LMPC) offers a compelling middle ground by using data from previous successful laps to refine the current trajectory. However, standard LMPC implementations—typically designed for 2D ground vehicles—fail in 3D drone racing due to mathematical singularities in progress tracking and a tendency to "cut corners" so aggressively that the vehicle eventually misses gates or crashes. This research addresses these hurdles to make LMPC a viable, safe, and real-time solution for high-speed flight.
Methodology
The proposed framework builds upon the standard LMPC architecture, which utilizes historical state and input trajectories to construct a "safe set" and a "cost-to-go" function. The authors introduce three specific innovations to adapt this for the aerial domain:
1. Cartesian-Based Formulation
Standard LMPC often relies on Frenet-frame dynamics to track progress along a path. In complex 3D racing tracks, Frenet frames can introduce integration errors and mathematical singularities. To avoid this, the authors utilize a 3D Cartesian corridor. They augment the drone's state with an arc-length variable to track progress. To ensure the system remains real-time capable (30Hz), a k-d tree is employed to efficiently search for the drone's current arc length based on its Cartesian position, with the estimation process taking only ~0.68 ms.
2. Adaptive Cost Function
Time-optimal controllers naturally seek the shortest path, which often leads to clipping gates. The authors developed an adaptive cost function that penalizes lateral deviation from the track centerline. This penalty is modulated by sigmoid functions, ensuring the penalty is highest exactly as the drone passes through a gate. This allows the drone to explore aggressive racing lines between gates while ensuring safe traversal through the gates themselves.
3. The Shifted Local Safe Set
A critical failure mode in standard LMPC is the "corner-cutting trap." As the controller optimizes for time, the historical safe set shifts toward the inside of turns. Eventually, the controller lacks reference states to steer the drone back toward the center, leading to a crash. The authors solve this by implementing a "Shifted Safe Set," which mathematically translates historical states to the opposite side of the central path. This provides the optimizer with a wider, more robust safety net, preventing the drone from becoming trapped on an unrecoverable trajectory.
Experimental Results
The algorithm was validated using a Crazyflie 2.1 quadrotor in both simulated and real-world environments.
- Performance Gains: When compared to a baseline PID controller, the LMPC reduced lap times by 60.85% (from 17.09s to 6.69s). More significantly, when initialized with MPCC++, a state-of-the-art model-based racing controller, the LMPC still managed to shave off an additional 6.05% of the lap time through iterative refinement.
- Real-Time Capability: The entire control loop, including the k-d tree search and the LMPC solver, runs comfortably at 30Hz on an offboard Intel i7 CPU, proving its feasibility for high-speed applications.
- Ablation Studies: Without the adaptive cost and shifted safe set, the drone consistently crashed or missed gates within 2 to 5 iterations. With these innovations, the system reliably converged to a safe, time-optimal path over 10 iterations.
Discussion and Limitations
While the results are promising, the authors identify several areas for future work. The algorithm currently converges to local optima, meaning the final performance is sensitive to the quality of the initial trajectory. Furthermore, to simplify the initial experiments, the drone's yaw angle was held constant. Future iterations will need to incorporate dynamic yaw control to reach the level of performance seen in professional human-piloted racing.
Despite these limitations, the research successfully demonstrates that with the right modifications for 3D safety and progress tracking, LMPC can bridge the gap between safe model-based control and the aggressive performance required for competitive autonomous racing.
Summary updated